Using voltage glitching to unleash the A100 in my mining card
Hackers are reverse‑engineering Nvidia’s CMP 170HX Ethereum mining cards to re‑enable disabled HBM memory and A100‑class features, in some cases turning nominal 8–10 GB cards into 64–80 GB AI accelerators. Commenters debate whether these fuses were blown mainly to salvage partially defective chips or to enforce artificial product segmentation and protect high‑margin datacenter SKUs, and note the economic incentives now driving such exploits as used card prices spike. The thread also touches on reliability risks, ECC error handling, cooling and power constraints, and broader questions about the ethics and practicality of selling deliberately hamstrung hardware.
Defects, binning, and product segmentation
- Several comments relate this to standard binning: same die used across tiers, with faulty parts fused off (like CPU cores).
- Others argue modern differences are driven more by product segmentation than by actual defects, especially late in a product’s life when yields are high.
- Some note that lower-tier SKUs can end up with flawless silicon that is deliberately disabled to fill demand.
HBM capacity, yields, and real defect rates
- Debate over whether CMP 170HX cards are limited due to bad HBM or just artificial nerfing.
- Some claim HBM yields are very good and that these mining cards were mostly binned for compute/logic failures, not memory.
- Anecdotes:
- “8GB→64GB” unlocks often very reliable; “10GB→80GB” unlocks more likely to hit bad RAM.
- Some A100-class cards remap faulty HBM rows internally; ECC can catch/mitigate errors.
Practicality and risks of the unlock
- Cooling is a major concern: cards lack proper coolers and expect loud server airflow; some users run them in mining/chassis setups.
- Power limiting is preferred over underclocking; performance loss is small.
- Faulty RAM can sometimes be “mapped out,” sacrificing only a few MB; others warn instability or silent corruption is possible without ECC.
Economics, ethics, and pricing of hamstrung hardware
- One side calls selling intentionally hamstrung hardware “morally repugnant.”
- Others argue segmentation is necessary:
- Makes lower-priced SKUs possible; without it only a single (higher) price point would exist.
- Using partially defective dies is seen as less wasteful.
- Disagreement over whether banning segmentation would benefit consumers; several argue it would mainly raise prices for budget buyers.
Reverse engineering, fuses, and security
- There has been a multi-year public reverse engineering effort on these GPUs, including academic work.
- Fuses here are one-time-programmable bits for configuration/segmentation, not safety devices.
- Some argue fuses should only ever restrict capabilities; others note they’re designed primarily for “economic extraction,” not security, so odd cases where blowing a fuse grants more access can exist.
- It’s expected such “mistakes” will be removed in future revisions.
Market impact and user reactions
- Prices of CMP 170HX have already spiked once unlocks became known.
- Multiple commenters express regret for not buying earlier or not holding longer; others recount profits flipping early cards.
- Consensus: unlocked cards offer A100-class AI capability (e.g., large local LLMs) at a fraction of datacenter pricing—if you acted early and can handle the quirks.